2 * Copyright (c) 2005 Voltaire Inc. All rights reserved.
3 * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
4 * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
5 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved.
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
21 * - Redistributions in binary form must reproduce the above
22 * copyright notice, this list of conditions and the following
23 * disclaimer in the documentation and/or other materials
24 * provided with the distribution.
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36 #include <linux/completion.h>
38 #include <linux/in6.h>
39 #include <linux/mutex.h>
40 #include <linux/random.h>
41 #include <linux/igmp.h>
42 #include <linux/idr.h>
43 #include <linux/inetdevice.h>
44 #include <linux/slab.h>
45 #include <linux/module.h>
46 #include <net/route.h>
48 #include <net/net_namespace.h>
49 #include <net/netns/generic.h>
52 #include <net/ip_fib.h>
53 #include <net/ip6_route.h>
55 #include <rdma/rdma_cm.h>
56 #include <rdma/rdma_cm_ib.h>
57 #include <rdma/rdma_netlink.h>
59 #include <rdma/ib_cache.h>
60 #include <rdma/ib_cm.h>
61 #include <rdma/ib_sa.h>
62 #include <rdma/iw_cm.h>
64 #include "core_priv.h"
66 MODULE_AUTHOR("Sean Hefty");
67 MODULE_DESCRIPTION("Generic RDMA CM Agent");
68 MODULE_LICENSE("Dual BSD/GPL");
70 #define CMA_CM_RESPONSE_TIMEOUT 20
71 #define CMA_MAX_CM_RETRIES 15
72 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
73 #define CMA_IBOE_PACKET_LIFETIME 18
75 static const char * const cma_events[] = {
76 [RDMA_CM_EVENT_ADDR_RESOLVED] = "address resolved",
77 [RDMA_CM_EVENT_ADDR_ERROR] = "address error",
78 [RDMA_CM_EVENT_ROUTE_RESOLVED] = "route resolved ",
79 [RDMA_CM_EVENT_ROUTE_ERROR] = "route error",
80 [RDMA_CM_EVENT_CONNECT_REQUEST] = "connect request",
81 [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
82 [RDMA_CM_EVENT_CONNECT_ERROR] = "connect error",
83 [RDMA_CM_EVENT_UNREACHABLE] = "unreachable",
84 [RDMA_CM_EVENT_REJECTED] = "rejected",
85 [RDMA_CM_EVENT_ESTABLISHED] = "established",
86 [RDMA_CM_EVENT_DISCONNECTED] = "disconnected",
87 [RDMA_CM_EVENT_DEVICE_REMOVAL] = "device removal",
88 [RDMA_CM_EVENT_MULTICAST_JOIN] = "multicast join",
89 [RDMA_CM_EVENT_MULTICAST_ERROR] = "multicast error",
90 [RDMA_CM_EVENT_ADDR_CHANGE] = "address change",
91 [RDMA_CM_EVENT_TIMEWAIT_EXIT] = "timewait exit",
94 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
98 return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
99 cma_events[index] : "unrecognized event";
101 EXPORT_SYMBOL(rdma_event_msg);
103 static void cma_add_one(struct ib_device *device);
104 static void cma_remove_one(struct ib_device *device, void *client_data);
106 static struct ib_client cma_client = {
109 .remove = cma_remove_one
112 static struct ib_sa_client sa_client;
113 static struct rdma_addr_client addr_client;
114 static LIST_HEAD(dev_list);
115 static LIST_HEAD(listen_any_list);
116 static DEFINE_MUTEX(lock);
117 static struct workqueue_struct *cma_wq;
118 static int cma_pernet_id;
127 static struct cma_pernet *cma_pernet(struct net *net)
129 return net_generic(net, cma_pernet_id);
132 static struct idr *cma_pernet_idr(struct net *net, enum rdma_port_space ps)
134 struct cma_pernet *pernet = cma_pernet(net);
138 return &pernet->tcp_ps;
140 return &pernet->udp_ps;
142 return &pernet->ipoib_ps;
144 return &pernet->ib_ps;
151 struct list_head list;
152 struct ib_device *device;
153 struct completion comp;
155 struct list_head id_list;
156 enum ib_gid_type *default_gid_type;
159 struct rdma_bind_list {
160 enum rdma_port_space ps;
161 struct hlist_head owners;
165 static int cma_ps_alloc(struct net *net, enum rdma_port_space ps,
166 struct rdma_bind_list *bind_list, int snum)
168 struct idr *idr = cma_pernet_idr(net, ps);
170 return idr_alloc(idr, bind_list, snum, snum + 1, GFP_KERNEL);
173 static struct rdma_bind_list *cma_ps_find(struct net *net,
174 enum rdma_port_space ps, int snum)
176 struct idr *idr = cma_pernet_idr(net, ps);
178 return idr_find(idr, snum);
181 static void cma_ps_remove(struct net *net, enum rdma_port_space ps, int snum)
183 struct idr *idr = cma_pernet_idr(net, ps);
185 idr_remove(idr, snum);
192 void cma_ref_dev(struct cma_device *cma_dev)
194 atomic_inc(&cma_dev->refcount);
197 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter filter,
200 struct cma_device *cma_dev;
201 struct cma_device *found_cma_dev = NULL;
205 list_for_each_entry(cma_dev, &dev_list, list)
206 if (filter(cma_dev->device, cookie)) {
207 found_cma_dev = cma_dev;
212 cma_ref_dev(found_cma_dev);
214 return found_cma_dev;
217 int cma_get_default_gid_type(struct cma_device *cma_dev,
220 if (port < rdma_start_port(cma_dev->device) ||
221 port > rdma_end_port(cma_dev->device))
224 return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
227 int cma_set_default_gid_type(struct cma_device *cma_dev,
229 enum ib_gid_type default_gid_type)
231 unsigned long supported_gids;
233 if (port < rdma_start_port(cma_dev->device) ||
234 port > rdma_end_port(cma_dev->device))
237 supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
239 if (!(supported_gids & 1 << default_gid_type))
242 cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
248 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
250 return cma_dev->device;
254 * Device removal can occur at anytime, so we need extra handling to
255 * serialize notifying the user of device removal with other callbacks.
256 * We do this by disabling removal notification while a callback is in process,
257 * and reporting it after the callback completes.
259 struct rdma_id_private {
260 struct rdma_cm_id id;
262 struct rdma_bind_list *bind_list;
263 struct hlist_node node;
264 struct list_head list; /* listen_any_list or cma_device.list */
265 struct list_head listen_list; /* per device listens */
266 struct cma_device *cma_dev;
267 struct list_head mc_list;
270 enum rdma_cm_state state;
272 struct mutex qp_mutex;
274 struct completion comp;
276 struct mutex handler_mutex;
280 struct ib_sa_query *query;
296 enum ib_gid_type gid_type;
299 struct cma_multicast {
300 struct rdma_id_private *id_priv;
302 struct ib_sa_multicast *ib;
304 struct list_head list;
306 struct sockaddr_storage addr;
312 struct work_struct work;
313 struct rdma_id_private *id;
314 enum rdma_cm_state old_state;
315 enum rdma_cm_state new_state;
316 struct rdma_cm_event event;
319 struct cma_ndev_work {
320 struct work_struct work;
321 struct rdma_id_private *id;
322 struct rdma_cm_event event;
325 struct iboe_mcast_work {
326 struct work_struct work;
327 struct rdma_id_private *id;
328 struct cma_multicast *mc;
341 u8 ip_version; /* IP version: 7:4 */
343 union cma_ip_addr src_addr;
344 union cma_ip_addr dst_addr;
347 #define CMA_VERSION 0x00
349 struct cma_req_info {
350 struct ib_device *device;
352 union ib_gid local_gid;
358 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
363 spin_lock_irqsave(&id_priv->lock, flags);
364 ret = (id_priv->state == comp);
365 spin_unlock_irqrestore(&id_priv->lock, flags);
369 static int cma_comp_exch(struct rdma_id_private *id_priv,
370 enum rdma_cm_state comp, enum rdma_cm_state exch)
375 spin_lock_irqsave(&id_priv->lock, flags);
376 if ((ret = (id_priv->state == comp)))
377 id_priv->state = exch;
378 spin_unlock_irqrestore(&id_priv->lock, flags);
382 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
383 enum rdma_cm_state exch)
386 enum rdma_cm_state old;
388 spin_lock_irqsave(&id_priv->lock, flags);
389 old = id_priv->state;
390 id_priv->state = exch;
391 spin_unlock_irqrestore(&id_priv->lock, flags);
395 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
397 return hdr->ip_version >> 4;
400 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
402 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
405 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
407 struct in_device *in_dev = NULL;
411 in_dev = __in_dev_get_rtnl(ndev);
414 ip_mc_inc_group(in_dev,
415 *(__be32 *)(mgid->raw + 12));
417 ip_mc_dec_group(in_dev,
418 *(__be32 *)(mgid->raw + 12));
422 return (in_dev) ? 0 : -ENODEV;
425 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
426 struct cma_device *cma_dev)
428 cma_ref_dev(cma_dev);
429 id_priv->cma_dev = cma_dev;
430 id_priv->gid_type = 0;
431 id_priv->id.device = cma_dev->device;
432 id_priv->id.route.addr.dev_addr.transport =
433 rdma_node_get_transport(cma_dev->device->node_type);
434 list_add_tail(&id_priv->list, &cma_dev->id_list);
437 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
438 struct cma_device *cma_dev)
440 _cma_attach_to_dev(id_priv, cma_dev);
442 cma_dev->default_gid_type[id_priv->id.port_num -
443 rdma_start_port(cma_dev->device)];
446 void cma_deref_dev(struct cma_device *cma_dev)
448 if (atomic_dec_and_test(&cma_dev->refcount))
449 complete(&cma_dev->comp);
452 static inline void release_mc(struct kref *kref)
454 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
456 kfree(mc->multicast.ib);
460 static void cma_release_dev(struct rdma_id_private *id_priv)
463 list_del(&id_priv->list);
464 cma_deref_dev(id_priv->cma_dev);
465 id_priv->cma_dev = NULL;
469 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
471 return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
474 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
476 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
479 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
481 return id_priv->id.route.addr.src_addr.ss_family;
484 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
486 struct ib_sa_mcmember_rec rec;
490 if (qkey && id_priv->qkey != qkey)
496 id_priv->qkey = qkey;
500 switch (id_priv->id.ps) {
503 id_priv->qkey = RDMA_UDP_QKEY;
506 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
507 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
508 id_priv->id.port_num, &rec.mgid,
511 id_priv->qkey = be32_to_cpu(rec.qkey);
519 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
521 dev_addr->dev_type = ARPHRD_INFINIBAND;
522 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
523 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
526 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
530 if (addr->sa_family != AF_IB) {
531 ret = rdma_translate_ip(addr, dev_addr, NULL);
533 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
540 static inline int cma_validate_port(struct ib_device *device, u8 port,
541 enum ib_gid_type gid_type,
542 union ib_gid *gid, int dev_type,
546 struct net_device *ndev = NULL;
548 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
551 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
554 if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
555 ndev = dev_get_by_index(&init_net, bound_if_index);
556 if (ndev && ndev->flags & IFF_LOOPBACK) {
557 pr_info("detected loopback device\n");
560 if (!device->get_netdev)
563 ndev = device->get_netdev(device, port);
568 gid_type = IB_GID_TYPE_IB;
571 ret = ib_find_cached_gid_by_port(device, gid, gid_type, port,
580 static int cma_acquire_dev(struct rdma_id_private *id_priv,
581 struct rdma_id_private *listen_id_priv)
583 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
584 struct cma_device *cma_dev;
585 union ib_gid gid, iboe_gid, *gidp;
589 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
590 id_priv->id.ps == RDMA_PS_IPOIB)
594 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
597 memcpy(&gid, dev_addr->src_dev_addr +
598 rdma_addr_gid_offset(dev_addr), sizeof gid);
600 if (listen_id_priv) {
601 cma_dev = listen_id_priv->cma_dev;
602 port = listen_id_priv->id.port_num;
603 gidp = rdma_protocol_roce(cma_dev->device, port) ?
606 ret = cma_validate_port(cma_dev->device, port,
607 rdma_protocol_ib(cma_dev->device, port) ?
609 listen_id_priv->gid_type, gidp,
611 dev_addr->bound_dev_if);
613 id_priv->id.port_num = port;
618 list_for_each_entry(cma_dev, &dev_list, list) {
619 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
620 if (listen_id_priv &&
621 listen_id_priv->cma_dev == cma_dev &&
622 listen_id_priv->id.port_num == port)
625 gidp = rdma_protocol_roce(cma_dev->device, port) ?
628 ret = cma_validate_port(cma_dev->device, port,
629 rdma_protocol_ib(cma_dev->device, port) ?
631 cma_dev->default_gid_type[port - 1],
632 gidp, dev_addr->dev_type,
633 dev_addr->bound_dev_if);
635 id_priv->id.port_num = port;
643 cma_attach_to_dev(id_priv, cma_dev);
650 * Select the source IB device and address to reach the destination IB address.
652 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
654 struct cma_device *cma_dev, *cur_dev;
655 struct sockaddr_ib *addr;
656 union ib_gid gid, sgid, *dgid;
662 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
663 dgid = (union ib_gid *) &addr->sib_addr;
664 pkey = ntohs(addr->sib_pkey);
666 list_for_each_entry(cur_dev, &dev_list, list) {
667 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
668 if (!rdma_cap_af_ib(cur_dev->device, p))
671 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
674 for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i,
677 if (!memcmp(&gid, dgid, sizeof(gid))) {
680 id_priv->id.port_num = p;
684 if (!cma_dev && (gid.global.subnet_prefix ==
685 dgid->global.subnet_prefix)) {
688 id_priv->id.port_num = p;
698 cma_attach_to_dev(id_priv, cma_dev);
699 addr = (struct sockaddr_ib *) cma_src_addr(id_priv);
700 memcpy(&addr->sib_addr, &sgid, sizeof sgid);
701 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
705 static void cma_deref_id(struct rdma_id_private *id_priv)
707 if (atomic_dec_and_test(&id_priv->refcount))
708 complete(&id_priv->comp);
711 static int cma_disable_callback(struct rdma_id_private *id_priv,
712 enum rdma_cm_state state)
714 mutex_lock(&id_priv->handler_mutex);
715 if (id_priv->state != state) {
716 mutex_unlock(&id_priv->handler_mutex);
722 struct rdma_cm_id *rdma_create_id(struct net *net,
723 rdma_cm_event_handler event_handler,
724 void *context, enum rdma_port_space ps,
725 enum ib_qp_type qp_type)
727 struct rdma_id_private *id_priv;
729 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
731 return ERR_PTR(-ENOMEM);
733 id_priv->owner = task_pid_nr(current);
734 id_priv->state = RDMA_CM_IDLE;
735 id_priv->id.context = context;
736 id_priv->id.event_handler = event_handler;
738 id_priv->id.qp_type = qp_type;
739 spin_lock_init(&id_priv->lock);
740 mutex_init(&id_priv->qp_mutex);
741 init_completion(&id_priv->comp);
742 atomic_set(&id_priv->refcount, 1);
743 mutex_init(&id_priv->handler_mutex);
744 INIT_LIST_HEAD(&id_priv->listen_list);
745 INIT_LIST_HEAD(&id_priv->mc_list);
746 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
747 id_priv->id.route.addr.dev_addr.net = get_net(net);
751 EXPORT_SYMBOL(rdma_create_id);
753 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
755 struct ib_qp_attr qp_attr;
756 int qp_attr_mask, ret;
758 qp_attr.qp_state = IB_QPS_INIT;
759 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
763 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
767 qp_attr.qp_state = IB_QPS_RTR;
768 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
772 qp_attr.qp_state = IB_QPS_RTS;
774 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
779 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
781 struct ib_qp_attr qp_attr;
782 int qp_attr_mask, ret;
784 qp_attr.qp_state = IB_QPS_INIT;
785 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
789 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
792 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
793 struct ib_qp_init_attr *qp_init_attr)
795 struct rdma_id_private *id_priv;
799 id_priv = container_of(id, struct rdma_id_private, id);
800 if (id->device != pd->device)
803 qp_init_attr->port_num = id->port_num;
804 qp = ib_create_qp(pd, qp_init_attr);
808 if (id->qp_type == IB_QPT_UD)
809 ret = cma_init_ud_qp(id_priv, qp);
811 ret = cma_init_conn_qp(id_priv, qp);
816 id_priv->qp_num = qp->qp_num;
817 id_priv->srq = (qp->srq != NULL);
823 EXPORT_SYMBOL(rdma_create_qp);
825 void rdma_destroy_qp(struct rdma_cm_id *id)
827 struct rdma_id_private *id_priv;
829 id_priv = container_of(id, struct rdma_id_private, id);
830 mutex_lock(&id_priv->qp_mutex);
831 ib_destroy_qp(id_priv->id.qp);
832 id_priv->id.qp = NULL;
833 mutex_unlock(&id_priv->qp_mutex);
835 EXPORT_SYMBOL(rdma_destroy_qp);
837 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
838 struct rdma_conn_param *conn_param)
840 struct ib_qp_attr qp_attr;
841 int qp_attr_mask, ret;
844 mutex_lock(&id_priv->qp_mutex);
845 if (!id_priv->id.qp) {
850 /* Need to update QP attributes from default values. */
851 qp_attr.qp_state = IB_QPS_INIT;
852 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
856 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
860 qp_attr.qp_state = IB_QPS_RTR;
861 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
865 ret = ib_query_gid(id_priv->id.device, id_priv->id.port_num,
866 qp_attr.ah_attr.grh.sgid_index, &sgid, NULL);
870 BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
873 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
874 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
876 mutex_unlock(&id_priv->qp_mutex);
880 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
881 struct rdma_conn_param *conn_param)
883 struct ib_qp_attr qp_attr;
884 int qp_attr_mask, ret;
886 mutex_lock(&id_priv->qp_mutex);
887 if (!id_priv->id.qp) {
892 qp_attr.qp_state = IB_QPS_RTS;
893 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
898 qp_attr.max_rd_atomic = conn_param->initiator_depth;
899 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
901 mutex_unlock(&id_priv->qp_mutex);
905 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
907 struct ib_qp_attr qp_attr;
910 mutex_lock(&id_priv->qp_mutex);
911 if (!id_priv->id.qp) {
916 qp_attr.qp_state = IB_QPS_ERR;
917 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
919 mutex_unlock(&id_priv->qp_mutex);
923 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
924 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
926 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
930 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
933 pkey = ib_addr_get_pkey(dev_addr);
935 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
936 pkey, &qp_attr->pkey_index);
940 qp_attr->port_num = id_priv->id.port_num;
941 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
943 if (id_priv->id.qp_type == IB_QPT_UD) {
944 ret = cma_set_qkey(id_priv, 0);
948 qp_attr->qkey = id_priv->qkey;
949 *qp_attr_mask |= IB_QP_QKEY;
951 qp_attr->qp_access_flags = 0;
952 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
957 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
960 struct rdma_id_private *id_priv;
963 id_priv = container_of(id, struct rdma_id_private, id);
964 if (rdma_cap_ib_cm(id->device, id->port_num)) {
965 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
966 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
968 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
971 if (qp_attr->qp_state == IB_QPS_RTR)
972 qp_attr->rq_psn = id_priv->seq_num;
973 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
974 if (!id_priv->cm_id.iw) {
975 qp_attr->qp_access_flags = 0;
976 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
978 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
985 EXPORT_SYMBOL(rdma_init_qp_attr);
987 static inline int cma_zero_addr(struct sockaddr *addr)
989 switch (addr->sa_family) {
991 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
993 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
995 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
1001 static inline int cma_loopback_addr(struct sockaddr *addr)
1003 switch (addr->sa_family) {
1005 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
1007 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
1009 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
1015 static inline int cma_any_addr(struct sockaddr *addr)
1017 return cma_zero_addr(addr) || cma_loopback_addr(addr);
1020 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
1022 if (src->sa_family != dst->sa_family)
1025 switch (src->sa_family) {
1027 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
1028 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
1030 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
1031 &((struct sockaddr_in6 *) dst)->sin6_addr);
1033 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1034 &((struct sockaddr_ib *) dst)->sib_addr);
1038 static __be16 cma_port(struct sockaddr *addr)
1040 struct sockaddr_ib *sib;
1042 switch (addr->sa_family) {
1044 return ((struct sockaddr_in *) addr)->sin_port;
1046 return ((struct sockaddr_in6 *) addr)->sin6_port;
1048 sib = (struct sockaddr_ib *) addr;
1049 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1050 be64_to_cpu(sib->sib_sid_mask)));
1056 static inline int cma_any_port(struct sockaddr *addr)
1058 return !cma_port(addr);
1061 static void cma_save_ib_info(struct sockaddr *src_addr,
1062 struct sockaddr *dst_addr,
1063 struct rdma_cm_id *listen_id,
1064 struct ib_sa_path_rec *path)
1066 struct sockaddr_ib *listen_ib, *ib;
1068 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1070 ib = (struct sockaddr_ib *)src_addr;
1071 ib->sib_family = AF_IB;
1073 ib->sib_pkey = path->pkey;
1074 ib->sib_flowinfo = path->flow_label;
1075 memcpy(&ib->sib_addr, &path->sgid, 16);
1076 ib->sib_sid = path->service_id;
1077 ib->sib_scope_id = 0;
1079 ib->sib_pkey = listen_ib->sib_pkey;
1080 ib->sib_flowinfo = listen_ib->sib_flowinfo;
1081 ib->sib_addr = listen_ib->sib_addr;
1082 ib->sib_sid = listen_ib->sib_sid;
1083 ib->sib_scope_id = listen_ib->sib_scope_id;
1085 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1088 ib = (struct sockaddr_ib *)dst_addr;
1089 ib->sib_family = AF_IB;
1091 ib->sib_pkey = path->pkey;
1092 ib->sib_flowinfo = path->flow_label;
1093 memcpy(&ib->sib_addr, &path->dgid, 16);
1098 static void cma_save_ip4_info(struct sockaddr *src_addr,
1099 struct sockaddr *dst_addr,
1100 struct cma_hdr *hdr,
1103 struct sockaddr_in *ip4;
1106 ip4 = (struct sockaddr_in *)src_addr;
1107 ip4->sin_family = AF_INET;
1108 ip4->sin_addr.s_addr = hdr->dst_addr.ip4.addr;
1109 ip4->sin_port = local_port;
1113 ip4 = (struct sockaddr_in *)dst_addr;
1114 ip4->sin_family = AF_INET;
1115 ip4->sin_addr.s_addr = hdr->src_addr.ip4.addr;
1116 ip4->sin_port = hdr->port;
1120 static void cma_save_ip6_info(struct sockaddr *src_addr,
1121 struct sockaddr *dst_addr,
1122 struct cma_hdr *hdr,
1125 struct sockaddr_in6 *ip6;
1128 ip6 = (struct sockaddr_in6 *)src_addr;
1129 ip6->sin6_family = AF_INET6;
1130 ip6->sin6_addr = hdr->dst_addr.ip6;
1131 ip6->sin6_port = local_port;
1135 ip6 = (struct sockaddr_in6 *)dst_addr;
1136 ip6->sin6_family = AF_INET6;
1137 ip6->sin6_addr = hdr->src_addr.ip6;
1138 ip6->sin6_port = hdr->port;
1142 static u16 cma_port_from_service_id(__be64 service_id)
1144 return (u16)be64_to_cpu(service_id);
1147 static int cma_save_ip_info(struct sockaddr *src_addr,
1148 struct sockaddr *dst_addr,
1149 struct ib_cm_event *ib_event,
1152 struct cma_hdr *hdr;
1155 hdr = ib_event->private_data;
1156 if (hdr->cma_version != CMA_VERSION)
1159 port = htons(cma_port_from_service_id(service_id));
1161 switch (cma_get_ip_ver(hdr)) {
1163 cma_save_ip4_info(src_addr, dst_addr, hdr, port);
1166 cma_save_ip6_info(src_addr, dst_addr, hdr, port);
1169 return -EAFNOSUPPORT;
1175 static int cma_save_net_info(struct sockaddr *src_addr,
1176 struct sockaddr *dst_addr,
1177 struct rdma_cm_id *listen_id,
1178 struct ib_cm_event *ib_event,
1179 sa_family_t sa_family, __be64 service_id)
1181 if (sa_family == AF_IB) {
1182 if (ib_event->event == IB_CM_REQ_RECEIVED)
1183 cma_save_ib_info(src_addr, dst_addr, listen_id,
1184 ib_event->param.req_rcvd.primary_path);
1185 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1186 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1190 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1193 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1194 struct cma_req_info *req)
1196 const struct ib_cm_req_event_param *req_param =
1197 &ib_event->param.req_rcvd;
1198 const struct ib_cm_sidr_req_event_param *sidr_param =
1199 &ib_event->param.sidr_req_rcvd;
1201 switch (ib_event->event) {
1202 case IB_CM_REQ_RECEIVED:
1203 req->device = req_param->listen_id->device;
1204 req->port = req_param->port;
1205 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1206 sizeof(req->local_gid));
1207 req->has_gid = true;
1208 req->service_id = req_param->primary_path->service_id;
1209 req->pkey = be16_to_cpu(req_param->primary_path->pkey);
1210 if (req->pkey != req_param->bth_pkey)
1211 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1212 "RDMA CMA: in the future this may cause the request to be dropped\n",
1213 req_param->bth_pkey, req->pkey);
1215 case IB_CM_SIDR_REQ_RECEIVED:
1216 req->device = sidr_param->listen_id->device;
1217 req->port = sidr_param->port;
1218 req->has_gid = false;
1219 req->service_id = sidr_param->service_id;
1220 req->pkey = sidr_param->pkey;
1221 if (req->pkey != sidr_param->bth_pkey)
1222 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1223 "RDMA CMA: in the future this may cause the request to be dropped\n",
1224 sidr_param->bth_pkey, req->pkey);
1233 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1234 const struct sockaddr_in *dst_addr,
1235 const struct sockaddr_in *src_addr)
1237 __be32 daddr = dst_addr->sin_addr.s_addr,
1238 saddr = src_addr->sin_addr.s_addr;
1239 struct fib_result res;
1244 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1245 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1246 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1247 ipv4_is_loopback(saddr))
1250 memset(&fl4, 0, sizeof(fl4));
1251 fl4.flowi4_iif = net_dev->ifindex;
1256 err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1257 ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1263 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1264 const struct sockaddr_in6 *dst_addr,
1265 const struct sockaddr_in6 *src_addr)
1267 #if IS_ENABLED(CONFIG_IPV6)
1268 const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1269 IPV6_ADDR_LINKLOCAL;
1270 struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1271 &src_addr->sin6_addr, net_dev->ifindex,
1278 ret = rt->rt6i_idev->dev == net_dev;
1287 static bool validate_net_dev(struct net_device *net_dev,
1288 const struct sockaddr *daddr,
1289 const struct sockaddr *saddr)
1291 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1292 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1293 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1294 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1296 switch (daddr->sa_family) {
1298 return saddr->sa_family == AF_INET &&
1299 validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1302 return saddr->sa_family == AF_INET6 &&
1303 validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1310 static struct net_device *cma_get_net_dev(struct ib_cm_event *ib_event,
1311 const struct cma_req_info *req)
1313 struct sockaddr_storage listen_addr_storage, src_addr_storage;
1314 struct sockaddr *listen_addr = (struct sockaddr *)&listen_addr_storage,
1315 *src_addr = (struct sockaddr *)&src_addr_storage;
1316 struct net_device *net_dev;
1317 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1320 err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1323 return ERR_PTR(err);
1325 net_dev = ib_get_net_dev_by_params(req->device, req->port, req->pkey,
1328 return ERR_PTR(-ENODEV);
1330 if (!validate_net_dev(net_dev, listen_addr, src_addr)) {
1332 return ERR_PTR(-EHOSTUNREACH);
1338 static enum rdma_port_space rdma_ps_from_service_id(__be64 service_id)
1340 return (be64_to_cpu(service_id) >> 16) & 0xffff;
1343 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1344 const struct cma_hdr *hdr)
1346 struct sockaddr *addr = cma_src_addr(id_priv);
1348 struct in6_addr ip6_addr;
1350 if (cma_any_addr(addr) && !id_priv->afonly)
1353 switch (addr->sa_family) {
1355 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1356 if (cma_get_ip_ver(hdr) != 4)
1358 if (!cma_any_addr(addr) &&
1359 hdr->dst_addr.ip4.addr != ip4_addr)
1363 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1364 if (cma_get_ip_ver(hdr) != 6)
1366 if (!cma_any_addr(addr) &&
1367 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1379 static bool cma_protocol_roce_dev_port(struct ib_device *device, int port_num)
1381 enum rdma_link_layer ll = rdma_port_get_link_layer(device, port_num);
1382 enum rdma_transport_type transport =
1383 rdma_node_get_transport(device->node_type);
1385 return ll == IB_LINK_LAYER_ETHERNET && transport == RDMA_TRANSPORT_IB;
1388 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1390 struct ib_device *device = id->device;
1391 const int port_num = id->port_num ?: rdma_start_port(device);
1393 return cma_protocol_roce_dev_port(device, port_num);
1396 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1397 const struct net_device *net_dev,
1400 const struct rdma_addr *addr = &id->route.addr;
1403 /* This request is an AF_IB request or a RoCE request */
1404 return (!id->port_num || id->port_num == port_num) &&
1405 (addr->src_addr.ss_family == AF_IB ||
1406 cma_protocol_roce_dev_port(id->device, port_num));
1408 return !addr->dev_addr.bound_dev_if ||
1409 (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1410 addr->dev_addr.bound_dev_if == net_dev->ifindex);
1413 static struct rdma_id_private *cma_find_listener(
1414 const struct rdma_bind_list *bind_list,
1415 const struct ib_cm_id *cm_id,
1416 const struct ib_cm_event *ib_event,
1417 const struct cma_req_info *req,
1418 const struct net_device *net_dev)
1420 struct rdma_id_private *id_priv, *id_priv_dev;
1423 return ERR_PTR(-EINVAL);
1425 hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1426 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1427 if (id_priv->id.device == cm_id->device &&
1428 cma_match_net_dev(&id_priv->id, net_dev, req->port))
1430 list_for_each_entry(id_priv_dev,
1431 &id_priv->listen_list,
1433 if (id_priv_dev->id.device == cm_id->device &&
1434 cma_match_net_dev(&id_priv_dev->id, net_dev, req->port))
1440 return ERR_PTR(-EINVAL);
1443 static struct rdma_id_private *cma_id_from_event(struct ib_cm_id *cm_id,
1444 struct ib_cm_event *ib_event,
1445 struct net_device **net_dev)
1447 struct cma_req_info req;
1448 struct rdma_bind_list *bind_list;
1449 struct rdma_id_private *id_priv;
1452 err = cma_save_req_info(ib_event, &req);
1454 return ERR_PTR(err);
1456 *net_dev = cma_get_net_dev(ib_event, &req);
1457 if (IS_ERR(*net_dev)) {
1458 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1459 /* Assuming the protocol is AF_IB */
1461 } else if (cma_protocol_roce_dev_port(req.device, req.port)) {
1462 /* TODO find the net dev matching the request parameters
1463 * through the RoCE GID table */
1466 return ERR_CAST(*net_dev);
1470 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1471 rdma_ps_from_service_id(req.service_id),
1472 cma_port_from_service_id(req.service_id));
1473 id_priv = cma_find_listener(bind_list, cm_id, ib_event, &req, *net_dev);
1474 if (IS_ERR(id_priv) && *net_dev) {
1482 static inline int cma_user_data_offset(struct rdma_id_private *id_priv)
1484 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1487 static void cma_cancel_route(struct rdma_id_private *id_priv)
1489 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1491 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1495 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1497 struct rdma_id_private *dev_id_priv;
1500 * Remove from listen_any_list to prevent added devices from spawning
1501 * additional listen requests.
1504 list_del(&id_priv->list);
1506 while (!list_empty(&id_priv->listen_list)) {
1507 dev_id_priv = list_entry(id_priv->listen_list.next,
1508 struct rdma_id_private, listen_list);
1509 /* sync with device removal to avoid duplicate destruction */
1510 list_del_init(&dev_id_priv->list);
1511 list_del(&dev_id_priv->listen_list);
1512 mutex_unlock(&lock);
1514 rdma_destroy_id(&dev_id_priv->id);
1517 mutex_unlock(&lock);
1520 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1521 enum rdma_cm_state state)
1524 case RDMA_CM_ADDR_QUERY:
1525 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1527 case RDMA_CM_ROUTE_QUERY:
1528 cma_cancel_route(id_priv);
1530 case RDMA_CM_LISTEN:
1531 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1532 cma_cancel_listens(id_priv);
1539 static void cma_release_port(struct rdma_id_private *id_priv)
1541 struct rdma_bind_list *bind_list = id_priv->bind_list;
1542 struct net *net = id_priv->id.route.addr.dev_addr.net;
1548 hlist_del(&id_priv->node);
1549 if (hlist_empty(&bind_list->owners)) {
1550 cma_ps_remove(net, bind_list->ps, bind_list->port);
1553 mutex_unlock(&lock);
1556 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1558 struct cma_multicast *mc;
1560 while (!list_empty(&id_priv->mc_list)) {
1561 mc = container_of(id_priv->mc_list.next,
1562 struct cma_multicast, list);
1563 list_del(&mc->list);
1564 if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1565 id_priv->id.port_num)) {
1566 ib_sa_free_multicast(mc->multicast.ib);
1569 if (mc->igmp_joined) {
1570 struct rdma_dev_addr *dev_addr =
1571 &id_priv->id.route.addr.dev_addr;
1572 struct net_device *ndev = NULL;
1574 if (dev_addr->bound_dev_if)
1575 ndev = dev_get_by_index(&init_net,
1576 dev_addr->bound_dev_if);
1579 &mc->multicast.ib->rec.mgid,
1584 kref_put(&mc->mcref, release_mc);
1589 void rdma_destroy_id(struct rdma_cm_id *id)
1591 struct rdma_id_private *id_priv;
1592 enum rdma_cm_state state;
1594 id_priv = container_of(id, struct rdma_id_private, id);
1595 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1596 cma_cancel_operation(id_priv, state);
1599 * Wait for any active callback to finish. New callbacks will find
1600 * the id_priv state set to destroying and abort.
1602 mutex_lock(&id_priv->handler_mutex);
1603 mutex_unlock(&id_priv->handler_mutex);
1605 if (id_priv->cma_dev) {
1606 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1607 if (id_priv->cm_id.ib)
1608 ib_destroy_cm_id(id_priv->cm_id.ib);
1609 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1610 if (id_priv->cm_id.iw)
1611 iw_destroy_cm_id(id_priv->cm_id.iw);
1613 cma_leave_mc_groups(id_priv);
1614 cma_release_dev(id_priv);
1617 cma_release_port(id_priv);
1618 cma_deref_id(id_priv);
1619 wait_for_completion(&id_priv->comp);
1621 if (id_priv->internal_id)
1622 cma_deref_id(id_priv->id.context);
1624 kfree(id_priv->id.route.path_rec);
1625 put_net(id_priv->id.route.addr.dev_addr.net);
1628 EXPORT_SYMBOL(rdma_destroy_id);
1630 static int cma_rep_recv(struct rdma_id_private *id_priv)
1634 ret = cma_modify_qp_rtr(id_priv, NULL);
1638 ret = cma_modify_qp_rts(id_priv, NULL);
1642 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1648 cma_modify_qp_err(id_priv);
1649 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1654 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1655 struct ib_cm_rep_event_param *rep_data,
1658 event->param.conn.private_data = private_data;
1659 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1660 event->param.conn.responder_resources = rep_data->responder_resources;
1661 event->param.conn.initiator_depth = rep_data->initiator_depth;
1662 event->param.conn.flow_control = rep_data->flow_control;
1663 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1664 event->param.conn.srq = rep_data->srq;
1665 event->param.conn.qp_num = rep_data->remote_qpn;
1668 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1670 struct rdma_id_private *id_priv = cm_id->context;
1671 struct rdma_cm_event event;
1674 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1675 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1676 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1677 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1680 memset(&event, 0, sizeof event);
1681 switch (ib_event->event) {
1682 case IB_CM_REQ_ERROR:
1683 case IB_CM_REP_ERROR:
1684 event.event = RDMA_CM_EVENT_UNREACHABLE;
1685 event.status = -ETIMEDOUT;
1687 case IB_CM_REP_RECEIVED:
1688 if (id_priv->id.qp) {
1689 event.status = cma_rep_recv(id_priv);
1690 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1691 RDMA_CM_EVENT_ESTABLISHED;
1693 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1695 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1696 ib_event->private_data);
1698 case IB_CM_RTU_RECEIVED:
1699 case IB_CM_USER_ESTABLISHED:
1700 event.event = RDMA_CM_EVENT_ESTABLISHED;
1702 case IB_CM_DREQ_ERROR:
1703 event.status = -ETIMEDOUT; /* fall through */
1704 case IB_CM_DREQ_RECEIVED:
1705 case IB_CM_DREP_RECEIVED:
1706 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1707 RDMA_CM_DISCONNECT))
1709 event.event = RDMA_CM_EVENT_DISCONNECTED;
1711 case IB_CM_TIMEWAIT_EXIT:
1712 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1714 case IB_CM_MRA_RECEIVED:
1717 case IB_CM_REJ_RECEIVED:
1718 cma_modify_qp_err(id_priv);
1719 event.status = ib_event->param.rej_rcvd.reason;
1720 event.event = RDMA_CM_EVENT_REJECTED;
1721 event.param.conn.private_data = ib_event->private_data;
1722 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1725 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
1730 ret = id_priv->id.event_handler(&id_priv->id, &event);
1732 /* Destroy the CM ID by returning a non-zero value. */
1733 id_priv->cm_id.ib = NULL;
1734 cma_exch(id_priv, RDMA_CM_DESTROYING);
1735 mutex_unlock(&id_priv->handler_mutex);
1736 rdma_destroy_id(&id_priv->id);
1740 mutex_unlock(&id_priv->handler_mutex);
1744 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1745 struct ib_cm_event *ib_event,
1746 struct net_device *net_dev)
1748 struct rdma_id_private *id_priv;
1749 struct rdma_cm_id *id;
1750 struct rdma_route *rt;
1751 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1752 const __be64 service_id =
1753 ib_event->param.req_rcvd.primary_path->service_id;
1756 id = rdma_create_id(listen_id->route.addr.dev_addr.net,
1757 listen_id->event_handler, listen_id->context,
1758 listen_id->ps, ib_event->param.req_rcvd.qp_type);
1762 id_priv = container_of(id, struct rdma_id_private, id);
1763 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1764 (struct sockaddr *)&id->route.addr.dst_addr,
1765 listen_id, ib_event, ss_family, service_id))
1769 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1770 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1775 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1776 if (rt->num_paths == 2)
1777 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1780 ret = rdma_copy_addr(&rt->addr.dev_addr, net_dev, NULL);
1784 if (!cma_protocol_roce(listen_id) &&
1785 cma_any_addr(cma_src_addr(id_priv))) {
1786 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1787 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1788 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1789 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
1790 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1795 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1797 id_priv->state = RDMA_CM_CONNECT;
1801 rdma_destroy_id(id);
1805 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1806 struct ib_cm_event *ib_event,
1807 struct net_device *net_dev)
1809 struct rdma_id_private *id_priv;
1810 struct rdma_cm_id *id;
1811 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1812 struct net *net = listen_id->route.addr.dev_addr.net;
1815 id = rdma_create_id(net, listen_id->event_handler, listen_id->context,
1816 listen_id->ps, IB_QPT_UD);
1820 id_priv = container_of(id, struct rdma_id_private, id);
1821 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1822 (struct sockaddr *)&id->route.addr.dst_addr,
1823 listen_id, ib_event, ss_family,
1824 ib_event->param.sidr_req_rcvd.service_id))
1828 ret = rdma_copy_addr(&id->route.addr.dev_addr, net_dev, NULL);
1832 if (!cma_any_addr(cma_src_addr(id_priv))) {
1833 ret = cma_translate_addr(cma_src_addr(id_priv),
1834 &id->route.addr.dev_addr);
1840 id_priv->state = RDMA_CM_CONNECT;
1843 rdma_destroy_id(id);
1847 static void cma_set_req_event_data(struct rdma_cm_event *event,
1848 struct ib_cm_req_event_param *req_data,
1849 void *private_data, int offset)
1851 event->param.conn.private_data = private_data + offset;
1852 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1853 event->param.conn.responder_resources = req_data->responder_resources;
1854 event->param.conn.initiator_depth = req_data->initiator_depth;
1855 event->param.conn.flow_control = req_data->flow_control;
1856 event->param.conn.retry_count = req_data->retry_count;
1857 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1858 event->param.conn.srq = req_data->srq;
1859 event->param.conn.qp_num = req_data->remote_qpn;
1862 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1864 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1865 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1866 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1867 (id->qp_type == IB_QPT_UD)) ||
1871 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1873 struct rdma_id_private *listen_id, *conn_id;
1874 struct rdma_cm_event event;
1875 struct net_device *net_dev;
1878 listen_id = cma_id_from_event(cm_id, ib_event, &net_dev);
1879 if (IS_ERR(listen_id))
1880 return PTR_ERR(listen_id);
1882 if (!cma_check_req_qp_type(&listen_id->id, ib_event)) {
1887 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN)) {
1888 ret = -ECONNABORTED;
1892 memset(&event, 0, sizeof event);
1893 offset = cma_user_data_offset(listen_id);
1894 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1895 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1896 conn_id = cma_new_udp_id(&listen_id->id, ib_event, net_dev);
1897 event.param.ud.private_data = ib_event->private_data + offset;
1898 event.param.ud.private_data_len =
1899 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1901 conn_id = cma_new_conn_id(&listen_id->id, ib_event, net_dev);
1902 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1903 ib_event->private_data, offset);
1910 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1911 ret = cma_acquire_dev(conn_id, listen_id);
1915 conn_id->cm_id.ib = cm_id;
1916 cm_id->context = conn_id;
1917 cm_id->cm_handler = cma_ib_handler;
1920 * Protect against the user destroying conn_id from another thread
1921 * until we're done accessing it.
1923 atomic_inc(&conn_id->refcount);
1924 ret = conn_id->id.event_handler(&conn_id->id, &event);
1928 * Acquire mutex to prevent user executing rdma_destroy_id()
1929 * while we're accessing the cm_id.
1932 if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
1933 (conn_id->id.qp_type != IB_QPT_UD))
1934 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1935 mutex_unlock(&lock);
1936 mutex_unlock(&conn_id->handler_mutex);
1937 mutex_unlock(&listen_id->handler_mutex);
1938 cma_deref_id(conn_id);
1944 cma_deref_id(conn_id);
1945 /* Destroy the CM ID by returning a non-zero value. */
1946 conn_id->cm_id.ib = NULL;
1948 cma_exch(conn_id, RDMA_CM_DESTROYING);
1949 mutex_unlock(&conn_id->handler_mutex);
1951 mutex_unlock(&listen_id->handler_mutex);
1953 rdma_destroy_id(&conn_id->id);
1962 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
1964 if (addr->sa_family == AF_IB)
1965 return ((struct sockaddr_ib *) addr)->sib_sid;
1967 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
1969 EXPORT_SYMBOL(rdma_get_service_id);
1971 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1973 struct rdma_id_private *id_priv = iw_id->context;
1974 struct rdma_cm_event event;
1976 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1977 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1979 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1982 memset(&event, 0, sizeof event);
1983 switch (iw_event->event) {
1984 case IW_CM_EVENT_CLOSE:
1985 event.event = RDMA_CM_EVENT_DISCONNECTED;
1987 case IW_CM_EVENT_CONNECT_REPLY:
1988 memcpy(cma_src_addr(id_priv), laddr,
1989 rdma_addr_size(laddr));
1990 memcpy(cma_dst_addr(id_priv), raddr,
1991 rdma_addr_size(raddr));
1992 switch (iw_event->status) {
1994 event.event = RDMA_CM_EVENT_ESTABLISHED;
1995 event.param.conn.initiator_depth = iw_event->ird;
1996 event.param.conn.responder_resources = iw_event->ord;
2000 event.event = RDMA_CM_EVENT_REJECTED;
2003 event.event = RDMA_CM_EVENT_UNREACHABLE;
2006 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2010 case IW_CM_EVENT_ESTABLISHED:
2011 event.event = RDMA_CM_EVENT_ESTABLISHED;
2012 event.param.conn.initiator_depth = iw_event->ird;
2013 event.param.conn.responder_resources = iw_event->ord;
2019 event.status = iw_event->status;
2020 event.param.conn.private_data = iw_event->private_data;
2021 event.param.conn.private_data_len = iw_event->private_data_len;
2022 ret = id_priv->id.event_handler(&id_priv->id, &event);
2024 /* Destroy the CM ID by returning a non-zero value. */
2025 id_priv->cm_id.iw = NULL;
2026 cma_exch(id_priv, RDMA_CM_DESTROYING);
2027 mutex_unlock(&id_priv->handler_mutex);
2028 rdma_destroy_id(&id_priv->id);
2032 mutex_unlock(&id_priv->handler_mutex);
2036 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2037 struct iw_cm_event *iw_event)
2039 struct rdma_cm_id *new_cm_id;
2040 struct rdma_id_private *listen_id, *conn_id;
2041 struct rdma_cm_event event;
2043 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2044 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2046 listen_id = cm_id->context;
2047 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
2048 return -ECONNABORTED;
2050 /* Create a new RDMA id for the new IW CM ID */
2051 new_cm_id = rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2052 listen_id->id.event_handler,
2053 listen_id->id.context,
2054 RDMA_PS_TCP, IB_QPT_RC);
2055 if (IS_ERR(new_cm_id)) {
2059 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2060 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2061 conn_id->state = RDMA_CM_CONNECT;
2063 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr, NULL);
2065 mutex_unlock(&conn_id->handler_mutex);
2066 rdma_destroy_id(new_cm_id);
2070 ret = cma_acquire_dev(conn_id, listen_id);
2072 mutex_unlock(&conn_id->handler_mutex);
2073 rdma_destroy_id(new_cm_id);
2077 conn_id->cm_id.iw = cm_id;
2078 cm_id->context = conn_id;
2079 cm_id->cm_handler = cma_iw_handler;
2081 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2082 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2084 memset(&event, 0, sizeof event);
2085 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2086 event.param.conn.private_data = iw_event->private_data;
2087 event.param.conn.private_data_len = iw_event->private_data_len;
2088 event.param.conn.initiator_depth = iw_event->ird;
2089 event.param.conn.responder_resources = iw_event->ord;
2092 * Protect against the user destroying conn_id from another thread
2093 * until we're done accessing it.
2095 atomic_inc(&conn_id->refcount);
2096 ret = conn_id->id.event_handler(&conn_id->id, &event);
2098 /* User wants to destroy the CM ID */
2099 conn_id->cm_id.iw = NULL;
2100 cma_exch(conn_id, RDMA_CM_DESTROYING);
2101 mutex_unlock(&conn_id->handler_mutex);
2102 cma_deref_id(conn_id);
2103 rdma_destroy_id(&conn_id->id);
2107 mutex_unlock(&conn_id->handler_mutex);
2108 cma_deref_id(conn_id);
2111 mutex_unlock(&listen_id->handler_mutex);
2115 static int cma_ib_listen(struct rdma_id_private *id_priv)
2117 struct sockaddr *addr;
2118 struct ib_cm_id *id;
2121 addr = cma_src_addr(id_priv);
2122 svc_id = rdma_get_service_id(&id_priv->id, addr);
2123 id = ib_cm_insert_listen(id_priv->id.device, cma_req_handler, svc_id);
2126 id_priv->cm_id.ib = id;
2131 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2134 struct iw_cm_id *id;
2136 id = iw_create_cm_id(id_priv->id.device,
2137 iw_conn_req_handler,
2142 id->tos = id_priv->tos;
2143 id_priv->cm_id.iw = id;
2145 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2146 rdma_addr_size(cma_src_addr(id_priv)));
2148 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2151 iw_destroy_cm_id(id_priv->cm_id.iw);
2152 id_priv->cm_id.iw = NULL;
2158 static int cma_listen_handler(struct rdma_cm_id *id,
2159 struct rdma_cm_event *event)
2161 struct rdma_id_private *id_priv = id->context;
2163 id->context = id_priv->id.context;
2164 id->event_handler = id_priv->id.event_handler;
2165 return id_priv->id.event_handler(id, event);
2168 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2169 struct cma_device *cma_dev)
2171 struct rdma_id_private *dev_id_priv;
2172 struct rdma_cm_id *id;
2173 struct net *net = id_priv->id.route.addr.dev_addr.net;
2176 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2179 id = rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2180 id_priv->id.qp_type);
2184 dev_id_priv = container_of(id, struct rdma_id_private, id);
2186 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2187 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2188 rdma_addr_size(cma_src_addr(id_priv)));
2190 _cma_attach_to_dev(dev_id_priv, cma_dev);
2191 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2192 atomic_inc(&id_priv->refcount);
2193 dev_id_priv->internal_id = 1;
2194 dev_id_priv->afonly = id_priv->afonly;
2196 ret = rdma_listen(id, id_priv->backlog);
2198 pr_warn("RDMA CMA: cma_listen_on_dev, error %d, listening on device %s\n",
2199 ret, cma_dev->device->name);
2202 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2204 struct cma_device *cma_dev;
2207 list_add_tail(&id_priv->list, &listen_any_list);
2208 list_for_each_entry(cma_dev, &dev_list, list)
2209 cma_listen_on_dev(id_priv, cma_dev);
2210 mutex_unlock(&lock);
2213 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2215 struct rdma_id_private *id_priv;
2217 id_priv = container_of(id, struct rdma_id_private, id);
2218 id_priv->tos = (u8) tos;
2220 EXPORT_SYMBOL(rdma_set_service_type);
2222 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
2225 struct cma_work *work = context;
2226 struct rdma_route *route;
2228 route = &work->id->id.route;
2231 route->num_paths = 1;
2232 *route->path_rec = *path_rec;
2234 work->old_state = RDMA_CM_ROUTE_QUERY;
2235 work->new_state = RDMA_CM_ADDR_RESOLVED;
2236 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2237 work->event.status = status;
2240 queue_work(cma_wq, &work->work);
2243 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
2244 struct cma_work *work)
2246 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2247 struct ib_sa_path_rec path_rec;
2248 ib_sa_comp_mask comp_mask;
2249 struct sockaddr_in6 *sin6;
2250 struct sockaddr_ib *sib;
2252 memset(&path_rec, 0, sizeof path_rec);
2253 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2254 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2255 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2256 path_rec.numb_path = 1;
2257 path_rec.reversible = 1;
2258 path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2260 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2261 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2262 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2264 switch (cma_family(id_priv)) {
2266 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2267 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2270 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2271 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2272 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2275 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2276 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2277 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2281 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2282 id_priv->id.port_num, &path_rec,
2283 comp_mask, timeout_ms,
2284 GFP_KERNEL, cma_query_handler,
2285 work, &id_priv->query);
2287 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2290 static void cma_work_handler(struct work_struct *_work)
2292 struct cma_work *work = container_of(_work, struct cma_work, work);
2293 struct rdma_id_private *id_priv = work->id;
2296 mutex_lock(&id_priv->handler_mutex);
2297 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2300 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2301 cma_exch(id_priv, RDMA_CM_DESTROYING);
2305 mutex_unlock(&id_priv->handler_mutex);
2306 cma_deref_id(id_priv);
2308 rdma_destroy_id(&id_priv->id);
2312 static void cma_ndev_work_handler(struct work_struct *_work)
2314 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2315 struct rdma_id_private *id_priv = work->id;
2318 mutex_lock(&id_priv->handler_mutex);
2319 if (id_priv->state == RDMA_CM_DESTROYING ||
2320 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2323 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2324 cma_exch(id_priv, RDMA_CM_DESTROYING);
2329 mutex_unlock(&id_priv->handler_mutex);
2330 cma_deref_id(id_priv);
2332 rdma_destroy_id(&id_priv->id);
2336 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
2338 struct rdma_route *route = &id_priv->id.route;
2339 struct cma_work *work;
2342 work = kzalloc(sizeof *work, GFP_KERNEL);
2347 INIT_WORK(&work->work, cma_work_handler);
2348 work->old_state = RDMA_CM_ROUTE_QUERY;
2349 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2350 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2352 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2353 if (!route->path_rec) {
2358 ret = cma_query_ib_route(id_priv, timeout_ms, work);
2364 kfree(route->path_rec);
2365 route->path_rec = NULL;
2371 int rdma_set_ib_paths(struct rdma_cm_id *id,
2372 struct ib_sa_path_rec *path_rec, int num_paths)
2374 struct rdma_id_private *id_priv;
2377 id_priv = container_of(id, struct rdma_id_private, id);
2378 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2379 RDMA_CM_ROUTE_RESOLVED))
2382 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
2384 if (!id->route.path_rec) {
2389 id->route.num_paths = num_paths;
2392 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2395 EXPORT_SYMBOL(rdma_set_ib_paths);
2397 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
2399 struct cma_work *work;
2401 work = kzalloc(sizeof *work, GFP_KERNEL);
2406 INIT_WORK(&work->work, cma_work_handler);
2407 work->old_state = RDMA_CM_ROUTE_QUERY;
2408 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2409 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2410 queue_work(cma_wq, &work->work);
2414 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2417 struct net_device *dev;
2419 prio = rt_tos2priority(tos);
2420 dev = ndev->priv_flags & IFF_802_1Q_VLAN ?
2421 vlan_dev_real_dev(ndev) : ndev;
2424 return netdev_get_prio_tc_map(dev, prio);
2426 #if IS_ENABLED(CONFIG_VLAN_8021Q)
2427 if (ndev->priv_flags & IFF_802_1Q_VLAN)
2428 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
2429 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2434 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2436 struct rdma_route *route = &id_priv->id.route;
2437 struct rdma_addr *addr = &route->addr;
2438 struct cma_work *work;
2440 struct net_device *ndev = NULL;
2443 work = kzalloc(sizeof *work, GFP_KERNEL);
2448 INIT_WORK(&work->work, cma_work_handler);
2450 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2451 if (!route->path_rec) {
2456 route->num_paths = 1;
2458 if (addr->dev_addr.bound_dev_if) {
2459 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
2463 if (ndev->flags & IFF_LOOPBACK) {
2465 if (!id_priv->id.device->get_netdev)
2468 ndev = id_priv->id.device->get_netdev(id_priv->id.device,
2469 id_priv->id.port_num);
2474 route->path_rec->net = &init_net;
2475 route->path_rec->ifindex = ndev->ifindex;
2476 route->path_rec->gid_type = id_priv->gid_type;
2483 memcpy(route->path_rec->dmac, addr->dev_addr.dst_dev_addr, ETH_ALEN);
2485 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2486 &route->path_rec->sgid);
2487 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2488 &route->path_rec->dgid);
2490 /* Use the hint from IP Stack to select GID Type */
2491 if (route->path_rec->gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2492 route->path_rec->gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2493 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2494 /* TODO: get the hoplimit from the inet/inet6 device */
2495 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2497 route->path_rec->hop_limit = 1;
2498 route->path_rec->reversible = 1;
2499 route->path_rec->pkey = cpu_to_be16(0xffff);
2500 route->path_rec->mtu_selector = IB_SA_EQ;
2501 route->path_rec->sl = iboe_tos_to_sl(ndev, id_priv->tos);
2502 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2503 route->path_rec->rate_selector = IB_SA_EQ;
2504 route->path_rec->rate = iboe_get_rate(ndev);
2506 route->path_rec->packet_life_time_selector = IB_SA_EQ;
2507 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
2508 if (!route->path_rec->mtu) {
2513 work->old_state = RDMA_CM_ROUTE_QUERY;
2514 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2515 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2516 work->event.status = 0;
2518 queue_work(cma_wq, &work->work);
2523 kfree(route->path_rec);
2524 route->path_rec = NULL;
2530 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
2532 struct rdma_id_private *id_priv;
2535 id_priv = container_of(id, struct rdma_id_private, id);
2536 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
2539 atomic_inc(&id_priv->refcount);
2540 if (rdma_cap_ib_sa(id->device, id->port_num))
2541 ret = cma_resolve_ib_route(id_priv, timeout_ms);
2542 else if (rdma_protocol_roce(id->device, id->port_num))
2543 ret = cma_resolve_iboe_route(id_priv);
2544 else if (rdma_protocol_iwarp(id->device, id->port_num))
2545 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2554 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2555 cma_deref_id(id_priv);
2558 EXPORT_SYMBOL(rdma_resolve_route);
2560 static void cma_set_loopback(struct sockaddr *addr)
2562 switch (addr->sa_family) {
2564 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2567 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2571 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2577 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2579 struct cma_device *cma_dev, *cur_dev;
2580 struct ib_port_attr port_attr;
2588 list_for_each_entry(cur_dev, &dev_list, list) {
2589 if (cma_family(id_priv) == AF_IB &&
2590 !rdma_cap_ib_cm(cur_dev->device, 1))
2596 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2597 if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2598 port_attr.state == IB_PORT_ACTIVE) {
2613 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid, NULL);
2617 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2621 id_priv->id.route.addr.dev_addr.dev_type =
2622 (rdma_protocol_ib(cma_dev->device, p)) ?
2623 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2625 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2626 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2627 id_priv->id.port_num = p;
2628 cma_attach_to_dev(id_priv, cma_dev);
2629 cma_set_loopback(cma_src_addr(id_priv));
2631 mutex_unlock(&lock);
2635 static void addr_handler(int status, struct sockaddr *src_addr,
2636 struct rdma_dev_addr *dev_addr, void *context)
2638 struct rdma_id_private *id_priv = context;
2639 struct rdma_cm_event event;
2641 memset(&event, 0, sizeof event);
2642 mutex_lock(&id_priv->handler_mutex);
2643 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2644 RDMA_CM_ADDR_RESOLVED))
2647 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2648 if (!status && !id_priv->cma_dev)
2649 status = cma_acquire_dev(id_priv, NULL);
2652 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2653 RDMA_CM_ADDR_BOUND))
2655 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2656 event.status = status;
2658 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2660 if (id_priv->id.event_handler(&id_priv->id, &event)) {
2661 cma_exch(id_priv, RDMA_CM_DESTROYING);
2662 mutex_unlock(&id_priv->handler_mutex);
2663 cma_deref_id(id_priv);
2664 rdma_destroy_id(&id_priv->id);
2668 mutex_unlock(&id_priv->handler_mutex);
2669 cma_deref_id(id_priv);
2672 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2674 struct cma_work *work;
2678 work = kzalloc(sizeof *work, GFP_KERNEL);
2682 if (!id_priv->cma_dev) {
2683 ret = cma_bind_loopback(id_priv);
2688 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2689 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2692 INIT_WORK(&work->work, cma_work_handler);
2693 work->old_state = RDMA_CM_ADDR_QUERY;
2694 work->new_state = RDMA_CM_ADDR_RESOLVED;
2695 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2696 queue_work(cma_wq, &work->work);
2703 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2705 struct cma_work *work;
2708 work = kzalloc(sizeof *work, GFP_KERNEL);
2712 if (!id_priv->cma_dev) {
2713 ret = cma_resolve_ib_dev(id_priv);
2718 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2719 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2722 INIT_WORK(&work->work, cma_work_handler);
2723 work->old_state = RDMA_CM_ADDR_QUERY;
2724 work->new_state = RDMA_CM_ADDR_RESOLVED;
2725 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2726 queue_work(cma_wq, &work->work);
2733 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2734 struct sockaddr *dst_addr)
2736 if (!src_addr || !src_addr->sa_family) {
2737 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2738 src_addr->sa_family = dst_addr->sa_family;
2739 if (dst_addr->sa_family == AF_INET6) {
2740 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
2741 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
2742 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
2743 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
2744 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
2745 } else if (dst_addr->sa_family == AF_IB) {
2746 ((struct sockaddr_ib *) src_addr)->sib_pkey =
2747 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2750 return rdma_bind_addr(id, src_addr);
2753 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2754 struct sockaddr *dst_addr, int timeout_ms)
2756 struct rdma_id_private *id_priv;
2759 id_priv = container_of(id, struct rdma_id_private, id);
2760 if (id_priv->state == RDMA_CM_IDLE) {
2761 ret = cma_bind_addr(id, src_addr, dst_addr);
2766 if (cma_family(id_priv) != dst_addr->sa_family)
2769 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2772 atomic_inc(&id_priv->refcount);
2773 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2774 if (cma_any_addr(dst_addr)) {
2775 ret = cma_resolve_loopback(id_priv);
2777 if (dst_addr->sa_family == AF_IB) {
2778 ret = cma_resolve_ib_addr(id_priv);
2780 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2781 dst_addr, &id->route.addr.dev_addr,
2782 timeout_ms, addr_handler, id_priv);
2790 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2791 cma_deref_id(id_priv);
2794 EXPORT_SYMBOL(rdma_resolve_addr);
2796 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2798 struct rdma_id_private *id_priv;
2799 unsigned long flags;
2802 id_priv = container_of(id, struct rdma_id_private, id);
2803 spin_lock_irqsave(&id_priv->lock, flags);
2804 if (reuse || id_priv->state == RDMA_CM_IDLE) {
2805 id_priv->reuseaddr = reuse;
2810 spin_unlock_irqrestore(&id_priv->lock, flags);
2813 EXPORT_SYMBOL(rdma_set_reuseaddr);
2815 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2817 struct rdma_id_private *id_priv;
2818 unsigned long flags;
2821 id_priv = container_of(id, struct rdma_id_private, id);
2822 spin_lock_irqsave(&id_priv->lock, flags);
2823 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2824 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2825 id_priv->afonly = afonly;
2830 spin_unlock_irqrestore(&id_priv->lock, flags);
2833 EXPORT_SYMBOL(rdma_set_afonly);
2835 static void cma_bind_port(struct rdma_bind_list *bind_list,
2836 struct rdma_id_private *id_priv)
2838 struct sockaddr *addr;
2839 struct sockaddr_ib *sib;
2843 addr = cma_src_addr(id_priv);
2844 port = htons(bind_list->port);
2846 switch (addr->sa_family) {
2848 ((struct sockaddr_in *) addr)->sin_port = port;
2851 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2854 sib = (struct sockaddr_ib *) addr;
2855 sid = be64_to_cpu(sib->sib_sid);
2856 mask = be64_to_cpu(sib->sib_sid_mask);
2857 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2858 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2861 id_priv->bind_list = bind_list;
2862 hlist_add_head(&id_priv->node, &bind_list->owners);
2865 static int cma_alloc_port(enum rdma_port_space ps,
2866 struct rdma_id_private *id_priv, unsigned short snum)
2868 struct rdma_bind_list *bind_list;
2871 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2875 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
2881 bind_list->port = (unsigned short)ret;
2882 cma_bind_port(bind_list, id_priv);
2886 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2889 static int cma_alloc_any_port(enum rdma_port_space ps,
2890 struct rdma_id_private *id_priv)
2892 static unsigned int last_used_port;
2893 int low, high, remaining;
2895 struct net *net = id_priv->id.route.addr.dev_addr.net;
2897 inet_get_local_port_range(net, &low, &high);
2898 remaining = (high - low) + 1;
2899 rover = prandom_u32() % remaining + low;
2901 if (last_used_port != rover &&
2902 !cma_ps_find(net, ps, (unsigned short)rover)) {
2903 int ret = cma_alloc_port(ps, id_priv, rover);
2905 * Remember previously used port number in order to avoid
2906 * re-using same port immediately after it is closed.
2909 last_used_port = rover;
2910 if (ret != -EADDRNOTAVAIL)
2915 if ((rover < low) || (rover > high))
2919 return -EADDRNOTAVAIL;
2923 * Check that the requested port is available. This is called when trying to
2924 * bind to a specific port, or when trying to listen on a bound port. In
2925 * the latter case, the provided id_priv may already be on the bind_list, but
2926 * we still need to check that it's okay to start listening.
2928 static int cma_check_port(struct rdma_bind_list *bind_list,
2929 struct rdma_id_private *id_priv, uint8_t reuseaddr)
2931 struct rdma_id_private *cur_id;
2932 struct sockaddr *addr, *cur_addr;
2934 addr = cma_src_addr(id_priv);
2935 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2936 if (id_priv == cur_id)
2939 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2943 cur_addr = cma_src_addr(cur_id);
2944 if (id_priv->afonly && cur_id->afonly &&
2945 (addr->sa_family != cur_addr->sa_family))
2948 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2949 return -EADDRNOTAVAIL;
2951 if (!cma_addr_cmp(addr, cur_addr))
2957 static int cma_use_port(enum rdma_port_space ps,
2958 struct rdma_id_private *id_priv)
2960 struct rdma_bind_list *bind_list;
2961 unsigned short snum;
2964 snum = ntohs(cma_port(cma_src_addr(id_priv)));
2965 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2968 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
2970 ret = cma_alloc_port(ps, id_priv, snum);
2972 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2974 cma_bind_port(bind_list, id_priv);
2979 static int cma_bind_listen(struct rdma_id_private *id_priv)
2981 struct rdma_bind_list *bind_list = id_priv->bind_list;
2985 if (bind_list->owners.first->next)
2986 ret = cma_check_port(bind_list, id_priv, 0);
2987 mutex_unlock(&lock);
2991 static enum rdma_port_space cma_select_inet_ps(
2992 struct rdma_id_private *id_priv)
2994 switch (id_priv->id.ps) {
2999 return id_priv->id.ps;
3006 static enum rdma_port_space cma_select_ib_ps(struct rdma_id_private *id_priv)
3008 enum rdma_port_space ps = 0;
3009 struct sockaddr_ib *sib;
3010 u64 sid_ps, mask, sid;
3012 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3013 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3014 sid = be64_to_cpu(sib->sib_sid) & mask;
3016 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3017 sid_ps = RDMA_IB_IP_PS_IB;
3019 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3020 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3021 sid_ps = RDMA_IB_IP_PS_TCP;
3023 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3024 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3025 sid_ps = RDMA_IB_IP_PS_UDP;
3030 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3031 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3032 be64_to_cpu(sib->sib_sid_mask));
3037 static int cma_get_port(struct rdma_id_private *id_priv)
3039 enum rdma_port_space ps;
3042 if (cma_family(id_priv) != AF_IB)
3043 ps = cma_select_inet_ps(id_priv);
3045 ps = cma_select_ib_ps(id_priv);
3047 return -EPROTONOSUPPORT;
3050 if (cma_any_port(cma_src_addr(id_priv)))
3051 ret = cma_alloc_any_port(ps, id_priv);
3053 ret = cma_use_port(ps, id_priv);
3054 mutex_unlock(&lock);
3059 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3060 struct sockaddr *addr)
3062 #if IS_ENABLED(CONFIG_IPV6)
3063 struct sockaddr_in6 *sin6;
3065 if (addr->sa_family != AF_INET6)
3068 sin6 = (struct sockaddr_in6 *) addr;
3070 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3073 if (!sin6->sin6_scope_id)
3076 dev_addr->bound_dev_if = sin6->sin6_scope_id;
3081 int rdma_listen(struct rdma_cm_id *id, int backlog)
3083 struct rdma_id_private *id_priv;
3086 id_priv = container_of(id, struct rdma_id_private, id);
3087 if (id_priv->state == RDMA_CM_IDLE) {
3088 id->route.addr.src_addr.ss_family = AF_INET;
3089 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3094 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
3097 if (id_priv->reuseaddr) {
3098 ret = cma_bind_listen(id_priv);
3103 id_priv->backlog = backlog;
3105 if (rdma_cap_ib_cm(id->device, 1)) {
3106 ret = cma_ib_listen(id_priv);
3109 } else if (rdma_cap_iw_cm(id->device, 1)) {
3110 ret = cma_iw_listen(id_priv, backlog);
3118 cma_listen_on_all(id_priv);
3122 id_priv->backlog = 0;
3123 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3126 EXPORT_SYMBOL(rdma_listen);
3128 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3130 struct rdma_id_private *id_priv;
3133 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3134 addr->sa_family != AF_IB)
3135 return -EAFNOSUPPORT;
3137 id_priv = container_of(id, struct rdma_id_private, id);
3138 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3141 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3145 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3146 if (!cma_any_addr(addr)) {
3147 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3151 ret = cma_acquire_dev(id_priv, NULL);
3156 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3157 if (addr->sa_family == AF_INET)
3158 id_priv->afonly = 1;
3159 #if IS_ENABLED(CONFIG_IPV6)
3160 else if (addr->sa_family == AF_INET6) {
3161 struct net *net = id_priv->id.route.addr.dev_addr.net;
3163 id_priv->afonly = net->ipv6.sysctl.bindv6only;
3167 ret = cma_get_port(id_priv);
3173 if (id_priv->cma_dev)
3174 cma_release_dev(id_priv);
3176 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3179 EXPORT_SYMBOL(rdma_bind_addr);
3181 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3183 struct cma_hdr *cma_hdr;
3186 cma_hdr->cma_version = CMA_VERSION;
3187 if (cma_family(id_priv) == AF_INET) {
3188 struct sockaddr_in *src4, *dst4;
3190 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3191 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3193 cma_set_ip_ver(cma_hdr, 4);
3194 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3195 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3196 cma_hdr->port = src4->sin_port;
3197 } else if (cma_family(id_priv) == AF_INET6) {
3198 struct sockaddr_in6 *src6, *dst6;
3200 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3201 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3203 cma_set_ip_ver(cma_hdr, 6);
3204 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3205 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3206 cma_hdr->port = src6->sin6_port;
3211 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3212 struct ib_cm_event *ib_event)
3214 struct rdma_id_private *id_priv = cm_id->context;
3215 struct rdma_cm_event event;
3216 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
3219 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
3222 memset(&event, 0, sizeof event);
3223 switch (ib_event->event) {
3224 case IB_CM_SIDR_REQ_ERROR:
3225 event.event = RDMA_CM_EVENT_UNREACHABLE;
3226 event.status = -ETIMEDOUT;
3228 case IB_CM_SIDR_REP_RECEIVED:
3229 event.param.ud.private_data = ib_event->private_data;
3230 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3231 if (rep->status != IB_SIDR_SUCCESS) {
3232 event.event = RDMA_CM_EVENT_UNREACHABLE;
3233 event.status = ib_event->param.sidr_rep_rcvd.status;
3236 ret = cma_set_qkey(id_priv, rep->qkey);
3238 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3242 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
3243 id_priv->id.route.path_rec,
3244 &event.param.ud.ah_attr);
3245 event.param.ud.qp_num = rep->qpn;
3246 event.param.ud.qkey = rep->qkey;
3247 event.event = RDMA_CM_EVENT_ESTABLISHED;
3251 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3256 ret = id_priv->id.event_handler(&id_priv->id, &event);
3258 /* Destroy the CM ID by returning a non-zero value. */
3259 id_priv->cm_id.ib = NULL;
3260 cma_exch(id_priv, RDMA_CM_DESTROYING);
3261 mutex_unlock(&id_priv->handler_mutex);
3262 rdma_destroy_id(&id_priv->id);
3266 mutex_unlock(&id_priv->handler_mutex);
3270 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3271 struct rdma_conn_param *conn_param)
3273 struct ib_cm_sidr_req_param req;
3274 struct ib_cm_id *id;
3278 memset(&req, 0, sizeof req);
3279 offset = cma_user_data_offset(id_priv);
3280 req.private_data_len = offset + conn_param->private_data_len;
3281 if (req.private_data_len < conn_param->private_data_len)
3284 if (req.private_data_len) {
3285 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3289 private_data = NULL;
3292 if (conn_param->private_data && conn_param->private_data_len)
3293 memcpy(private_data + offset, conn_param->private_data,
3294 conn_param->private_data_len);
3297 ret = cma_format_hdr(private_data, id_priv);
3300 req.private_data = private_data;
3303 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3309 id_priv->cm_id.ib = id;
3311 req.path = id_priv->id.route.path_rec;
3312 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3313 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3314 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3316 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3318 ib_destroy_cm_id(id_priv->cm_id.ib);
3319 id_priv->cm_id.ib = NULL;
3322 kfree(private_data);
3326 static int cma_connect_ib(struct rdma_id_private *id_priv,
3327 struct rdma_conn_param *conn_param)
3329 struct ib_cm_req_param req;
3330 struct rdma_route *route;
3332 struct ib_cm_id *id;
3335 memset(&req, 0, sizeof req);
3336 offset = cma_user_data_offset(id_priv);
3337 req.private_data_len = offset + conn_param->private_data_len;
3338 if (req.private_data_len < conn_param->private_data_len)
3341 if (req.private_data_len) {
3342 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3346 private_data = NULL;
3349 if (conn_param->private_data && conn_param->private_data_len)
3350 memcpy(private_data + offset, conn_param->private_data,
3351 conn_param->private_data_len);
3353 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3358 id_priv->cm_id.ib = id;
3360 route = &id_priv->id.route;
3362 ret = cma_format_hdr(private_data, id_priv);
3365 req.private_data = private_data;
3368 req.primary_path = &route->path_rec[0];
3369 if (route->num_paths == 2)
3370 req.alternate_path = &route->path_rec[1];
3372 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3373 req.qp_num = id_priv->qp_num;
3374 req.qp_type = id_priv->id.qp_type;
3375 req.starting_psn = id_priv->seq_num;
3376 req.responder_resources = conn_param->responder_resources;
3377 req.initiator_depth = conn_param->initiator_depth;
3378 req.flow_control = conn_param->flow_control;
3379 req.retry_count = min_t(u8, 7, conn_param->retry_count);
3380 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3381 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3382 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3383 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3384 req.srq = id_priv->srq ? 1 : 0;
3386 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3388 if (ret && !IS_ERR(id)) {
3389 ib_destroy_cm_id(id);
3390 id_priv->cm_id.ib = NULL;
3393 kfree(private_data);
3397 static int cma_connect_iw(struct rdma_id_private *id_priv,
3398 struct rdma_conn_param *conn_param)
3400 struct iw_cm_id *cm_id;
3402 struct iw_cm_conn_param iw_param;
3404 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3406 return PTR_ERR(cm_id);
3408 cm_id->tos = id_priv->tos;
3409 id_priv->cm_id.iw = cm_id;
3411 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3412 rdma_addr_size(cma_src_addr(id_priv)));
3413 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3414 rdma_addr_size(cma_dst_addr(id_priv)));
3416 ret = cma_modify_qp_rtr(id_priv, conn_param);
3421 iw_param.ord = conn_param->initiator_depth;
3422 iw_param.ird = conn_param->responder_resources;
3423 iw_param.private_data = conn_param->private_data;
3424 iw_param.private_data_len = conn_param->private_data_len;
3425 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3427 memset(&iw_param, 0, sizeof iw_param);
3428 iw_param.qpn = id_priv->qp_num;
3430 ret = iw_cm_connect(cm_id, &iw_param);
3433 iw_destroy_cm_id(cm_id);
3434 id_priv->cm_id.iw = NULL;
3439 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3441 struct rdma_id_private *id_priv;
3444 id_priv = container_of(id, struct rdma_id_private, id);
3445 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
3449 id_priv->qp_num = conn_param->qp_num;
3450 id_priv->srq = conn_param->srq;
3453 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3454 if (id->qp_type == IB_QPT_UD)
3455 ret = cma_resolve_ib_udp(id_priv, conn_param);
3457 ret = cma_connect_ib(id_priv, conn_param);
3458 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3459 ret = cma_connect_iw(id_priv, conn_param);
3467 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
3470 EXPORT_SYMBOL(rdma_connect);
3472 static int cma_accept_ib(struct rdma_id_private *id_priv,
3473 struct rdma_conn_param *conn_param)
3475 struct ib_cm_rep_param rep;
3478 ret = cma_modify_qp_rtr(id_priv, conn_param);
3482 ret = cma_modify_qp_rts(id_priv, conn_param);
3486 memset(&rep, 0, sizeof rep);
3487 rep.qp_num = id_priv->qp_num;
3488 rep.starting_psn = id_priv->seq_num;
3489 rep.private_data = conn_param->private_data;
3490 rep.private_data_len = conn_param->private_data_len;
3491 rep.responder_resources = conn_param->responder_resources;
3492 rep.initiator_depth = conn_param->initiator_depth;
3493 rep.failover_accepted = 0;
3494 rep.flow_control = conn_param->flow_control;
3495 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3496 rep.srq = id_priv->srq ? 1 : 0;
3498 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
3503 static int cma_accept_iw(struct rdma_id_private *id_priv,
3504 struct rdma_conn_param *conn_param)
3506 struct iw_cm_conn_param iw_param;
3509 ret = cma_modify_qp_rtr(id_priv, conn_param);
3513 iw_param.ord = conn_param->initiator_depth;
3514 iw_param.ird = conn_param->responder_resources;
3515 iw_param.private_data = conn_param->private_data;
3516 iw_param.private_data_len = conn_param->private_data_len;
3517 if (id_priv->id.qp) {
3518 iw_param.qpn = id_priv->qp_num;
3520 iw_param.qpn = conn_param->qp_num;
3522 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3525 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3526 enum ib_cm_sidr_status status, u32 qkey,
3527 const void *private_data, int private_data_len)
3529 struct ib_cm_sidr_rep_param rep;
3532 memset(&rep, 0, sizeof rep);
3533 rep.status = status;
3534 if (status == IB_SIDR_SUCCESS) {
3535 ret = cma_set_qkey(id_priv, qkey);
3538 rep.qp_num = id_priv->qp_num;
3539 rep.qkey = id_priv->qkey;
3541 rep.private_data = private_data;
3542 rep.private_data_len = private_data_len;
3544 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3547 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3549 struct rdma_id_private *id_priv;
3552 id_priv = container_of(id, struct rdma_id_private, id);
3554 id_priv->owner = task_pid_nr(current);
3556 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3559 if (!id->qp && conn_param) {
3560 id_priv->qp_num = conn_param->qp_num;
3561 id_priv->srq = conn_param->srq;
3564 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3565 if (id->qp_type == IB_QPT_UD) {
3567 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3569 conn_param->private_data,
3570 conn_param->private_data_len);
3572 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3576 ret = cma_accept_ib(id_priv, conn_param);
3578 ret = cma_rep_recv(id_priv);
3580 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3581 ret = cma_accept_iw(id_priv, conn_param);
3590 cma_modify_qp_err(id_priv);
3591 rdma_reject(id, NULL, 0);
3594 EXPORT_SYMBOL(rdma_accept);
3596 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3598 struct rdma_id_private *id_priv;
3601 id_priv = container_of(id, struct rdma_id_private, id);
3602 if (!id_priv->cm_id.ib)
3605 switch (id->device->node_type) {
3606 case RDMA_NODE_IB_CA:
3607 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3615 EXPORT_SYMBOL(rdma_notify);
3617 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3618 u8 private_data_len)
3620 struct rdma_id_private *id_priv;
3623 id_priv = container_of(id, struct rdma_id_private, id);
3624 if (!id_priv->cm_id.ib)
3627 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3628 if (id->qp_type == IB_QPT_UD)
3629 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3630 private_data, private_data_len);
3632 ret = ib_send_cm_rej(id_priv->cm_id.ib,
3633 IB_CM_REJ_CONSUMER_DEFINED, NULL,
3634 0, private_data, private_data_len);
3635 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3636 ret = iw_cm_reject(id_priv->cm_id.iw,
3637 private_data, private_data_len);
3643 EXPORT_SYMBOL(rdma_reject);
3645 int rdma_disconnect(struct rdma_cm_id *id)
3647 struct rdma_id_private *id_priv;
3650 id_priv = container_of(id, struct rdma_id_private, id);
3651 if (!id_priv->cm_id.ib)
3654 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3655 ret = cma_modify_qp_err(id_priv);
3658 /* Initiate or respond to a disconnect. */
3659 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3660 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3661 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3662 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3669 EXPORT_SYMBOL(rdma_disconnect);
3671 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3673 struct rdma_id_private *id_priv;
3674 struct cma_multicast *mc = multicast->context;
3675 struct rdma_cm_event event;
3678 id_priv = mc->id_priv;
3679 if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
3680 cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
3684 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3685 mutex_lock(&id_priv->qp_mutex);
3686 if (!status && id_priv->id.qp)
3687 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3688 be16_to_cpu(multicast->rec.mlid));
3689 mutex_unlock(&id_priv->qp_mutex);
3691 memset(&event, 0, sizeof event);
3692 event.status = status;
3693 event.param.ud.private_data = mc->context;
3695 struct rdma_dev_addr *dev_addr =
3696 &id_priv->id.route.addr.dev_addr;
3697 struct net_device *ndev =
3698 dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3699 enum ib_gid_type gid_type =
3700 id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
3701 rdma_start_port(id_priv->cma_dev->device)];
3703 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3704 ib_init_ah_from_mcmember(id_priv->id.device,
3705 id_priv->id.port_num, &multicast->rec,
3707 &event.param.ud.ah_attr);
3708 event.param.ud.qp_num = 0xFFFFFF;
3709 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3713 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3715 ret = id_priv->id.event_handler(&id_priv->id, &event);
3717 cma_exch(id_priv, RDMA_CM_DESTROYING);
3718 mutex_unlock(&id_priv->handler_mutex);
3719 rdma_destroy_id(&id_priv->id);
3723 mutex_unlock(&id_priv->handler_mutex);
3727 static void cma_set_mgid(struct rdma_id_private *id_priv,
3728 struct sockaddr *addr, union ib_gid *mgid)
3730 unsigned char mc_map[MAX_ADDR_LEN];
3731 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3732 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3733 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3735 if (cma_any_addr(addr)) {
3736 memset(mgid, 0, sizeof *mgid);
3737 } else if ((addr->sa_family == AF_INET6) &&
3738 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3740 /* IPv6 address is an SA assigned MGID. */
3741 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3742 } else if (addr->sa_family == AF_IB) {
3743 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3744 } else if ((addr->sa_family == AF_INET6)) {
3745 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3746 if (id_priv->id.ps == RDMA_PS_UDP)
3747 mc_map[7] = 0x01; /* Use RDMA CM signature */
3748 *mgid = *(union ib_gid *) (mc_map + 4);
3750 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3751 if (id_priv->id.ps == RDMA_PS_UDP)
3752 mc_map[7] = 0x01; /* Use RDMA CM signature */
3753 *mgid = *(union ib_gid *) (mc_map + 4);
3757 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3758 struct cma_multicast *mc)
3760 struct ib_sa_mcmember_rec rec;
3761 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3762 ib_sa_comp_mask comp_mask;
3765 ib_addr_get_mgid(dev_addr, &rec.mgid);
3766 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3771 ret = cma_set_qkey(id_priv, 0);
3775 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3776 rec.qkey = cpu_to_be32(id_priv->qkey);
3777 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3778 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3781 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3782 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3783 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3784 IB_SA_MCMEMBER_REC_FLOW_LABEL |
3785 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3787 if (id_priv->id.ps == RDMA_PS_IPOIB)
3788 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3789 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3790 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3791 IB_SA_MCMEMBER_REC_MTU |
3792 IB_SA_MCMEMBER_REC_HOP_LIMIT;
3794 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3795 id_priv->id.port_num, &rec,
3796 comp_mask, GFP_KERNEL,
3797 cma_ib_mc_handler, mc);
3798 return PTR_ERR_OR_ZERO(mc->multicast.ib);
3801 static void iboe_mcast_work_handler(struct work_struct *work)
3803 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3804 struct cma_multicast *mc = mw->mc;
3805 struct ib_sa_multicast *m = mc->multicast.ib;
3807 mc->multicast.ib->context = mc;
3808 cma_ib_mc_handler(0, m);
3809 kref_put(&mc->mcref, release_mc);
3813 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3815 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3816 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3818 if (cma_any_addr(addr)) {
3819 memset(mgid, 0, sizeof *mgid);
3820 } else if (addr->sa_family == AF_INET6) {
3821 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3823 mgid->raw[0] = 0xff;
3824 mgid->raw[1] = 0x0e;
3833 mgid->raw[10] = 0xff;
3834 mgid->raw[11] = 0xff;
3835 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3839 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3840 struct cma_multicast *mc)
3842 struct iboe_mcast_work *work;
3843 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3845 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3846 struct net_device *ndev = NULL;
3847 enum ib_gid_type gid_type;
3849 if (cma_zero_addr((struct sockaddr *)&mc->addr))
3852 work = kzalloc(sizeof *work, GFP_KERNEL);
3856 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3857 if (!mc->multicast.ib) {
3862 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3864 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3865 if (id_priv->id.ps == RDMA_PS_UDP)
3866 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3868 if (dev_addr->bound_dev_if)
3869 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3874 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3875 mc->multicast.ib->rec.hop_limit = 1;
3876 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3878 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
3879 rdma_start_port(id_priv->cma_dev->device)];
3880 if (addr->sa_family == AF_INET) {
3881 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
3882 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid,
3885 mc->igmp_joined = true;
3886 mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
3889 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
3893 if (err || !mc->multicast.ib->rec.mtu) {
3898 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3899 &mc->multicast.ib->rec.port_gid);
3902 INIT_WORK(&work->work, iboe_mcast_work_handler);
3903 kref_get(&mc->mcref);
3904 queue_work(cma_wq, &work->work);
3909 kfree(mc->multicast.ib);
3915 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3918 struct rdma_id_private *id_priv;
3919 struct cma_multicast *mc;
3922 id_priv = container_of(id, struct rdma_id_private, id);
3923 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3924 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3927 mc = kmalloc(sizeof *mc, GFP_KERNEL);
3931 memcpy(&mc->addr, addr, rdma_addr_size(addr));
3932 mc->context = context;
3933 mc->id_priv = id_priv;
3934 mc->igmp_joined = false;
3935 spin_lock(&id_priv->lock);
3936 list_add(&mc->list, &id_priv->mc_list);
3937 spin_unlock(&id_priv->lock);
3939 if (rdma_protocol_roce(id->device, id->port_num)) {
3940 kref_init(&mc->mcref);
3941 ret = cma_iboe_join_multicast(id_priv, mc);
3942 } else if (rdma_cap_ib_mcast(id->device, id->port_num))
3943 ret = cma_join_ib_multicast(id_priv, mc);
3948 spin_lock_irq(&id_priv->lock);
3949 list_del(&mc->list);
3950 spin_unlock_irq(&id_priv->lock);
3955 EXPORT_SYMBOL(rdma_join_multicast);
3957 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3959 struct rdma_id_private *id_priv;
3960 struct cma_multicast *mc;
3962 id_priv = container_of(id, struct rdma_id_private, id);
3963 spin_lock_irq(&id_priv->lock);
3964 list_for_each_entry(mc, &id_priv->mc_list, list) {
3965 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
3966 list_del(&mc->list);
3967 spin_unlock_irq(&id_priv->lock);
3970 ib_detach_mcast(id->qp,
3971 &mc->multicast.ib->rec.mgid,
3972 be16_to_cpu(mc->multicast.ib->rec.mlid));
3974 BUG_ON(id_priv->cma_dev->device != id->device);
3976 if (rdma_cap_ib_mcast(id->device, id->port_num)) {
3977 ib_sa_free_multicast(mc->multicast.ib);
3979 } else if (rdma_protocol_roce(id->device, id->port_num)) {
3980 if (mc->igmp_joined) {
3981 struct rdma_dev_addr *dev_addr =
3982 &id->route.addr.dev_addr;
3983 struct net_device *ndev = NULL;
3985 if (dev_addr->bound_dev_if)
3986 ndev = dev_get_by_index(&init_net,
3987 dev_addr->bound_dev_if);
3990 &mc->multicast.ib->rec.mgid,
3994 mc->igmp_joined = false;
3996 kref_put(&mc->mcref, release_mc);
4001 spin_unlock_irq(&id_priv->lock);
4003 EXPORT_SYMBOL(rdma_leave_multicast);
4005 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4007 struct rdma_dev_addr *dev_addr;
4008 struct cma_ndev_work *work;
4010 dev_addr = &id_priv->id.route.addr.dev_addr;
4012 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4013 (net_eq(dev_net(ndev), dev_addr->net)) &&
4014 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4015 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4016 ndev->name, &id_priv->id);
4017 work = kzalloc(sizeof *work, GFP_KERNEL);
4021 INIT_WORK(&work->work, cma_ndev_work_handler);
4023 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4024 atomic_inc(&id_priv->refcount);
4025 queue_work(cma_wq, &work->work);
4031 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4034 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4035 struct cma_device *cma_dev;
4036 struct rdma_id_private *id_priv;
4037 int ret = NOTIFY_DONE;
4039 if (event != NETDEV_BONDING_FAILOVER)
4042 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
4046 list_for_each_entry(cma_dev, &dev_list, list)
4047 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4048 ret = cma_netdev_change(ndev, id_priv);
4054 mutex_unlock(&lock);
4058 static struct notifier_block cma_nb = {
4059 .notifier_call = cma_netdev_callback
4062 static void cma_add_one(struct ib_device *device)
4064 struct cma_device *cma_dev;
4065 struct rdma_id_private *id_priv;
4067 unsigned long supported_gids = 0;
4069 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4073 cma_dev->device = device;
4074 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4075 sizeof(*cma_dev->default_gid_type),
4077 if (!cma_dev->default_gid_type) {
4081 for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
4082 supported_gids = roce_gid_type_mask_support(device, i);
4083 WARN_ON(!supported_gids);
4084 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4085 find_first_bit(&supported_gids, BITS_PER_LONG);
4088 init_completion(&cma_dev->comp);
4089 atomic_set(&cma_dev->refcount, 1);
4090 INIT_LIST_HEAD(&cma_dev->id_list);
4091 ib_set_client_data(device, &cma_client, cma_dev);
4094 list_add_tail(&cma_dev->list, &dev_list);
4095 list_for_each_entry(id_priv, &listen_any_list, list)
4096 cma_listen_on_dev(id_priv, cma_dev);
4097 mutex_unlock(&lock);
4100 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
4102 struct rdma_cm_event event;
4103 enum rdma_cm_state state;
4106 /* Record that we want to remove the device */
4107 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
4108 if (state == RDMA_CM_DESTROYING)
4111 cma_cancel_operation(id_priv, state);
4112 mutex_lock(&id_priv->handler_mutex);
4114 /* Check for destruction from another callback. */
4115 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
4118 memset(&event, 0, sizeof event);
4119 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
4120 ret = id_priv->id.event_handler(&id_priv->id, &event);
4122 mutex_unlock(&id_priv->handler_mutex);
4126 static void cma_process_remove(struct cma_device *cma_dev)
4128 struct rdma_id_private *id_priv;
4132 while (!list_empty(&cma_dev->id_list)) {
4133 id_priv = list_entry(cma_dev->id_list.next,
4134 struct rdma_id_private, list);
4136 list_del(&id_priv->listen_list);
4137 list_del_init(&id_priv->list);
4138 atomic_inc(&id_priv->refcount);
4139 mutex_unlock(&lock);
4141 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
4142 cma_deref_id(id_priv);
4144 rdma_destroy_id(&id_priv->id);
4148 mutex_unlock(&lock);
4150 cma_deref_dev(cma_dev);
4151 wait_for_completion(&cma_dev->comp);
4154 static void cma_remove_one(struct ib_device *device, void *client_data)
4156 struct cma_device *cma_dev = client_data;
4162 list_del(&cma_dev->list);
4163 mutex_unlock(&lock);
4165 cma_process_remove(cma_dev);
4166 kfree(cma_dev->default_gid_type);
4170 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
4172 struct nlmsghdr *nlh;
4173 struct rdma_cm_id_stats *id_stats;
4174 struct rdma_id_private *id_priv;
4175 struct rdma_cm_id *id = NULL;
4176 struct cma_device *cma_dev;
4177 int i_dev = 0, i_id = 0;
4180 * We export all of the IDs as a sequence of messages. Each
4181 * ID gets its own netlink message.
4185 list_for_each_entry(cma_dev, &dev_list, list) {
4186 if (i_dev < cb->args[0]) {
4192 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4193 if (i_id < cb->args[1]) {
4198 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
4199 sizeof *id_stats, RDMA_NL_RDMA_CM,
4200 RDMA_NL_RDMA_CM_ID_STATS,
4205 memset(id_stats, 0, sizeof *id_stats);
4207 id_stats->node_type = id->route.addr.dev_addr.dev_type;
4208 id_stats->port_num = id->port_num;
4209 id_stats->bound_dev_if =
4210 id->route.addr.dev_addr.bound_dev_if;
4212 if (ibnl_put_attr(skb, nlh,
4213 rdma_addr_size(cma_src_addr(id_priv)),
4214 cma_src_addr(id_priv),
4215 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
4217 if (ibnl_put_attr(skb, nlh,
4218 rdma_addr_size(cma_src_addr(id_priv)),
4219 cma_dst_addr(id_priv),
4220 RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
4223 id_stats->pid = id_priv->owner;
4224 id_stats->port_space = id->ps;
4225 id_stats->cm_state = id_priv->state;
4226 id_stats->qp_num = id_priv->qp_num;
4227 id_stats->qp_type = id->qp_type;
4237 mutex_unlock(&lock);
4238 cb->args[0] = i_dev;
4244 static const struct ibnl_client_cbs cma_cb_table[] = {
4245 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
4246 .module = THIS_MODULE },
4249 static int cma_init_net(struct net *net)
4251 struct cma_pernet *pernet = cma_pernet(net);
4253 idr_init(&pernet->tcp_ps);
4254 idr_init(&pernet->udp_ps);
4255 idr_init(&pernet->ipoib_ps);
4256 idr_init(&pernet->ib_ps);
4261 static void cma_exit_net(struct net *net)
4263 struct cma_pernet *pernet = cma_pernet(net);
4265 idr_destroy(&pernet->tcp_ps);
4266 idr_destroy(&pernet->udp_ps);
4267 idr_destroy(&pernet->ipoib_ps);
4268 idr_destroy(&pernet->ib_ps);
4271 static struct pernet_operations cma_pernet_operations = {
4272 .init = cma_init_net,
4273 .exit = cma_exit_net,
4274 .id = &cma_pernet_id,
4275 .size = sizeof(struct cma_pernet),
4278 static int __init cma_init(void)
4282 cma_wq = create_singlethread_workqueue("rdma_cm");
4286 ret = register_pernet_subsys(&cma_pernet_operations);
4290 ib_sa_register_client(&sa_client);
4291 rdma_addr_register_client(&addr_client);
4292 register_netdevice_notifier(&cma_nb);
4294 ret = ib_register_client(&cma_client);
4298 if (ibnl_add_client(RDMA_NL_RDMA_CM, ARRAY_SIZE(cma_cb_table),
4300 pr_warn("RDMA CMA: failed to add netlink callback\n");
4301 cma_configfs_init();
4306 unregister_netdevice_notifier(&cma_nb);
4307 rdma_addr_unregister_client(&addr_client);
4308 ib_sa_unregister_client(&sa_client);
4310 destroy_workqueue(cma_wq);
4314 static void __exit cma_cleanup(void)
4316 cma_configfs_exit();
4317 ibnl_remove_client(RDMA_NL_RDMA_CM);
4318 ib_unregister_client(&cma_client);
4319 unregister_netdevice_notifier(&cma_nb);
4320 rdma_addr_unregister_client(&addr_client);
4321 ib_sa_unregister_client(&sa_client);
4322 unregister_pernet_subsys(&cma_pernet_operations);
4323 destroy_workqueue(cma_wq);
4326 module_init(cma_init);
4327 module_exit(cma_cleanup);